Complex routing problems are critical in achieving operational efficiency. These problems are typically addressed sequentially using cluster-first, route-second frameworks. Unfortunately, this method can lead to suboptimal results due to a myopic clustering phase. To overcome this limitation, we propose a strategy that incorporates information about the optimal tour lengths of potential clusters early in the process. This adjustment aims to make the framework less myopic. We introduce rapid and highly accurate estimators of Traveling Salesperson Problem (TSP) tour lengths based on neural networks (NNs). Our study is distinct in that we train our models using newly designed instances that closely replicate real-world logistics networks and morphologies. These instances introduce significant computational challenges, which we address using an unsupervised learning method to derive tight lower bounds and partial TSP solutions, subsequently used as predictors. By combining theoretical knowledge in the routing domain with neural networks, our approach achieves predictions with an average deviation of less than 0.7\% from optimality. Extensive benchmarking shows that our models exhibit up to 100 times lower prediction errors than existing machine learning (ML) approaches on out-of-distribution test instances. Moreover, we integrate our models into a metaheuristic framework, creating an enumeration-like solution framework, capable of efficiently solving routing problems involving up to 100,000 cities. Compared to the current state-of-the-art solvers, our method significantly improves both solution time and quality, demonstrating the effectiveness of our features, models, and methodology.

Transfer speed and protection are critical factors that influence passengers' willingness to use public transportation. Due to unpredictable traffic patterns, fixed transfer schedules may not always align. This work proposes two online stochastic optimization algorithms for the transfer synchronization problem using the hold, skip-stop, and speedup tactics. First, we design an offline arc-flow model using time-expanded graphs to enumerate all possible tactics. The model minimizes total passenger travel time by reducing transfer times. Then we implement two online stochastic optimization algorithms based on the offline model in a discrete-event dynamic environment. Decisions are made based on predictions of the future state of the bus network using sampling to generate scenarios from historical and real-time data. The performance of the algorithms is compared using a real dataset from the public transit system of Laval, Canada. The results show significant improvements in both the number of successful transfers and total passenger travel time across 29 bus lines. This supports the practicality of using online stochastic optimization algorithms to solve the transfer synchronization problem in real-world transportation systems.

This paper focuses on facility location problems with facility disruption, uncertain demand, and transportation costs without capacity constraints. Firstly, we analyzed the sources of uncertainty information in the facility location problems and the adaptability and tractability of the distributionally robust model under different uncertain information. Secondly, we incorporate the probability of disaster occurrence and the effects of disruption propagation into the construction of an ambiguity set for the disruption scenario. Furthermore, we introduce information regarding support, mean, and upper-bound distribution into the construction of ambiguity sets for uncertain demand and transportation cost scenarios. In terms of tractability, for all the mentioned scenarios, we reformulate the distributional robust model as a mixed-integer linear program. Finally, we validate the effectiveness and applicability of our model through case studies involving emergency facility placement in earthquake scenarios and the positioning of epidemic detection facilities.